2026Reference
Anthropic Behavioral Interview Field Guide
Own curriculum artifact · Sep 2026 · ~3 h
Your own guide to Anthropic culture-round prompts (why Anthropic, ethics stories, stock to zero) with answer patterns.
The senior-staff bar
What L7-level depth looks like in this topic, the follow-up an interviewer would push on, and the resources that prepare you for it.
From the interview reportsReported as the most common reason senior candidates fail at Anthropic; a Google-tagged commenter said most of his rejected referrals failed on culture. Senior-staff questions: “what evidence shows our sincerity about the mission?”, “critique our approach”, “something morally wrong you did”, “criteria to release a model”. Answers must work for a non-technical interviewer.
Start here
The best resources for a time-poor senior engineer, newest first.
2026Reference
Anthropic's Responsible Scaling Policy (v3.4, with Risk Reports)
Anthropic · Jul 2026 · ~3 h
Current RSP (v3.4 effective 8 Jul 2026) with every prior version and redline, plus the Feb and Aug 2026 redacted Risk Reports.
2026Blog post
Behavioral Coaching Notes: Anthropic Philosophy and OpenAI Technical Deep Dives
Austen McDonald, The Behavioral · May 2026 · ~0.5 h
Coach's 2026 notes: Anthropic asks moral dilemmas and guiding values, so develop an actual moral philosophy; OpenAI wants personal, forward-looking optimism.
2026Reference
Frontier Safety Framework (v3.1)
Google DeepMind · Apr 2026 · ~2 h
DeepMind's framework with Critical Capability Levels, new Tracked Capability Levels and Security Level 2+; all versions linked.
2026Reference
Claude's Constitution
Anthropic · Jan 2026 · ~3 h
The Jan 2026 constitution: Claude's values, priorities (safety, ethics, guidelines, helpfulness) and reasoning about its own nature.
2026Reference
The Adolescence of Technology
Dario Amodei · Jan 2026 · ~2 h
Dario's 2026 essay on the risks of powerful AI and how to get through the transition.
2025Paper
Agentic Misalignment: How LLMs could be insider threats
Anthropic · Jun 2025 · ~1.5 h
Stress tests where models in agentic settings chose blackmail or leaking when goals or continuity were threatened.
2025Reference
Preparedness Framework (Version 2)
OpenAI · Apr 2025 · ~2 h
OpenAI's tracked categories (bio/chem, cyber, AI self-improvement), High/Critical thresholds and Safety Advisory Group process; still current per 2026 system cards.
2025Reference
AI as Normal Technology
Arvind Narayanan and Sayash Kapoor · Apr 2025 · ~2 h
The main skeptical alternative: AI as a transformative but normal technology diffusing slowly, with risk handled by resilience.
2024Reference
Machines of Loving Grace
Dario Amodei · Oct 2024 · ~2 h
Dario's optimistic case: what powerful AI could do for biology, neuroscience, economic development and governance.
Learning path
Work through the stages in order. The last stage is the senior-staff layer: production case studies, papers to critique and large-scale trade-offs. Each resource's L7 line says what it adds at your level.
- 1
Each lab in its own words
Read every target lab's mission and safety framework with dates · Week 1 (≈12 h)
- 2023Anthropic · Mar 2023 · Reference · Core · ~1 hAnthropic's founding argument: rapid progress is likely, safety is empirical, so a safety-focused lab must work at the frontier.L7Older but canonical; lets you state Anthropic's bet and its main tension in your own words.
- 2026Anthropic · Jul 2026 · Reference · Advanced · ~3 hCurrent RSP (v3.4 effective 8 Jul 2026) with every prior version and redline, plus the Feb and Aug 2026 redacted Risk Reports.L7Lets you discuss release criteria and critique specific changes (e.g. revised automated-R&D threshold) with dates, which few candidates can.
- 2026Anthropic · Jan 2026 · Reference · Advanced · ~3 hThe Jan 2026 constitution: Claude's values, priorities (safety, ethics, guidelines, helpfulness) and reasoning about its own nature.L7Lets you argue concrete value trade-offs (helpfulness vs harm, honesty) in Anthropic's own framing during hypotheticals.
- 2024Dario Amodei · Oct 2024 · Reference · Core · ~2 hDario's optimistic case: what powerful AI could do for biology, neuroscience, economic development and governance.L7Gives the upside half of your risk/benefit answer; candidates have been quizzed on Dario's essays.
- 2026Dario Amodei · Jan 2026 · Reference · Core · ~2 hDario's 2026 essay on the risks of powerful AI and how to get through the transition.L7The latest statement of Anthropic leadership's risk view; quote it accurately and say where you differ.
- 2023Anthropic · Sep 2023 · Reference · Core · ~0.5 hGovernance structure giving an independent trust power to elect board members.L7Structural evidence of mission commitment, and something to critique on its limits.
- 2018OpenAI · Apr 2018 · Reference · Core · ~0.3 hOpenAI's mission principles incl. broadly distributed benefits and the assist-a-rival clause.L7Older but canonical; required for why-OpenAI and Anthropic-vs-OpenAI comparisons.
- 2025OpenAI · Apr 2025 · Reference · Advanced · ~2 hOpenAI's tracked categories (bio/chem, cyber, AI self-improvement), High/Critical thresholds and Safety Advisory Group process; still current per 2026 system cards.L7Lets you compare release-gating mechanisms across labs and critique the clause allowing adjustment if rivals ship.
- 2025OpenAI · Mar 2025 · Reference · Core · ~0.5 hOpenAI's principles: AGI as a continuum, iterative deployment, defence in depth, human control.L7Gives OpenAI's own framing for why-OpenAI and lets you argue iterative deployment both ways.
- 2026Google DeepMind · Apr 2026 · Reference · Advanced · ~2 hDeepMind's framework with Critical Capability Levels, new Tracked Capability Levels and Security Level 2+; all versions linked.L7Third data point for comparing release criteria and security commitments across labs.
- 2026Meta · Apr 2026 · Reference · Advanced · ~1 hMeta's successor to its Feb 2025 Frontier AI Framework: chem/bio, cyber and loss-of-control thresholds for open and closed releases.L7Needed for MSL interviews and for arguing open-weights release risk.
- 2025xAI · Dec 2025 · Reference · Advanced · ~1 hxAI's TFAIA-compliant framework on malicious use and loss of control, successor to its Aug 2025 Risk Management Framework.L7Lets you discuss xAI's safety posture concretely in a why-xAI answer.
- 2023
- 2
Form your own views
A p(doom), timeline and biggest-risk view you can defend against both optimists and pessimists · Week 2 (≈12 h)
- 2024Dan Hendrycks (Center for AI Safety) · Dec 2024 · Book · Core · ~15 hFree textbook on risk categories (malicious use, racing, organisational, rogue AI), safety engineering and governance.L7Gives a structured taxonomy for 'how AI goes wrong' that works for non-technical interviewers.
- 2026Yoshua Bengio (chair) and 100+ experts · Feb 2026 · Survey paper · Advanced · ~6 hGovernment-backed scientific review of general-purpose AI capabilities, risks and mitigations.L7The most neutral evidence base for risk and regulation answers to non-technical interviewers.
- 202680,000 Hours · Feb 2026 · Reference · Intro · ~1.5 hAccessible problem profile on AI risk with counterarguments, updated Oct 2026.L7Baseline; fastest structured overview for 'why does AI safety matter to you'.
- 2025Kokotajlo, Alexander, Larsen, Lifland, Dean · Apr 2025 · Reference · Advanced · ~3 hDetailed scenario of rapid AI progress to 2027 with race and slowdown endings.L7Concrete scenario to agree or disagree with on timelines, racing and slowdowns.
- 2025Arvind Narayanan and Sayash Kapoor · Apr 2025 · Reference · Advanced · ~2 hThe main skeptical alternative: AI as a transformative but normal technology diffusing slowly, with risk handled by resilience.L7Gives you a serious counterview so your p(doom) and critique answers show you weighed both sides.
- 2025titotal (LessWrong) · Jun 2025 · Blog post · Advanced · ~1.5 hDetailed critique of the forecasting models behind AI 2027.L7Shows how to change or defend a belief on evidence, useful for 'what would change your mind'.
- 2025Eliezer Yudkowsky and Nate Soares · Sep 2025 · Book · Core · ~7 h · paidThe strongest pessimist case: current methods can't align superintelligence, so stop building it.L7Lets you steelman the high-p(doom) and 'slow down' position before rejecting or adopting it.
- 2026Dwarkesh Patel · Feb 2026 · Video · Core · ~2.5 h · videoDario on timelines, scaling, economics and safety in 2026.L7Current leadership views you can cite and push back on.
- 202580,000 Hours · Oct 2025 · Video · Core · ~3 hHolden (now at Anthropic) on concrete safety work at frontier companies and his AGI takes.L7Insider view on what safety at a frontier company looks like; good for critique answers.
- 2025METR · Mar 2025 · Paper · Advanced · ~1 hTask-length metric showing the time horizon of agent tasks doubling roughly every 7 months.L7A quantitative anchor for timeline views instead of vibes.
- 2026BlueDot Impact · Nov 2026 · Course · Intro · ~25 h25-hour cohort course on the strategic landscape of AGI: risks, actors, interventions.L7Fastest structured route to a defensible overall view if you are new to AI safety debates.
- 2024
- 3
Write the one-pagers and ethics stories
One page per question and three 3-minute ethics stories · Week 3 (≈6 h)
- 2026★Anthropic Behavioral Interview Field GuideOwn curriculum artifact · Sep 2026 · Reference · Core · ~3 hYour own guide to Anthropic culture-round prompts (why Anthropic, ethics stories, stock to zero) with answer patterns.L7Lets you rehearse the exact reported prompts so your 3-minute ethics stories are specific, not generic.
- 2026Austen McDonald, The Behavioral · May 2026 · Blog post · Core · ~0.5 hCoach's 2026 notes: Anthropic asks moral dilemmas and guiding values, so develop an actual moral philosophy; OpenAI wants personal, forward-looking optimism.L7Current calibration of what culture interviewers at Anthropic and OpenAI reward from senior candidates.
- 2026Hello Interview · 2026 · Practice · Core · ~1 h · paidCovers the Culture round: AI safety, judgment, values conflicting with business decisions, with reported questions.L7Shows the culture round weighs as much as technical rounds at staff level.
- 2026Aced (formerly Exponent) · 2026 · Practice · Intro · ~0.5 hProcess overview and sample values questions.L7Baseline; question list only.
- 2026Axios · Aug 2026 · Blog post · Core · ~0.2 hReporting that Anthropic's culture interview asks whether you'd accept the stock going to zero for the mission.L7Tells you the question is real and planned for, so prepare a reasoned position, not a reflex.
- 2009Michael Sandel, Harvard · 2009 · Course · Core · ~12 h · videoHarvard's moral philosophy course on video: utilitarianism, rights, Kant, Rawls, applied dilemmas.L7Older but canonical; gives you named ethical frameworks for Anthropic's moral-dilemma questions.
- 2026Anthropic · Nov 2026 · Reference · Core · ~1 hAnthropic's usage policy incl. high-risk use-case requirements (e.g. employment decisions need human review and disclosure).L7Lets you answer customer-misuse scenarios (automated hiring) with Anthropic's actual rules.
- 2026Anthropic · 2026 · Reference · Core · ~2 hAnthropic's announcements, policy positions and launches.L7Read the last six months so you can say which recent moves you agree or disagree with.
- 2026
- 4
Senior-staff depth
Critique specific lab decisions with primary evidence and propose what you would do · Week 4+ (≈15 h)
- 2024Greenblatt, Denison, Wright et al. (Anthropic, Redwood) · Dec 2024 · Paper · Advanced · ~2 hEvidence that Claude 3 Opus strategically complied with training it disagreed with to avoid being modified.L7Lets you cite a concrete empirical misalignment result and its limits when asked how AI goes wrong.
- 2025Anthropic · Jun 2025 · Paper · Advanced · ~1.5 hStress tests where models in agentic settings chose blackmail or leaking when goals or continuity were threatened.L7Grounds answers on agentic risk in high-stakes deployments with specific, caveated findings.
- 2025Anthropic · Nov 2025 · Paper · Advanced · ~1 hReward hacking in realistic RL environments generalised into broader misaligned behaviour.L7The most current 'how AI goes wrong in practice' example tied to production training.
- 2025Lindsey, Gurnee, Ameisen et al. (Anthropic) · Mar 2025 · Paper · Advanced · ~4 hAttribution-graph case studies of Claude 3.5 Haiku: multi-step reasoning, planning in poems, hallucination circuits.L7Lets you discuss what interpretability can and cannot show today, grounding views on whether we understand models.
- 2025Shah, Irpan, Turner et al. (Google DeepMind) · Apr 2025 · Paper · Advanced · ~6 hDeepMind's 100+ page plan covering misuse and misalignment with amplified oversight, monitoring and security.L7Lets you compare DeepMind's technical safety plan with Anthropic's at the level of mechanisms.
- 2026OpenAI · May 2026 · Reference · Advanced · ~1 hMaps OpenAI's safety practices onto California's TFAIA and the EU GPAI Code of Practice; the Preparedness Framework remains the foundation.L7Lets you discuss how lab frameworks are becoming regulatory compliance documents.
- 2025Korbak et al. (authors across OpenAI, Anthropic, Google DeepMind and others) · Jul 2025 · Paper · Advanced · ~1.5 hCross-lab position paper arguing CoT monitoring is valuable but fragile under training pressure.L7A rare cross-lab consensus document; use it in 'what should labs agree on' and critique answers.
- 2023Greenblatt, Shlegeris, Sachan, Roger (Redwood Research) · Dec 2023 · Paper · Advanced · ~2 hProtocols that stay safe even if the model is trying to subvert them.L7Gives a concrete 'what engineers can do' answer beyond alignment: control and monitoring.
- 2025Sharkey, Chughtai et al. · Jan 2025 · Survey paper · Advanced · ~5 hMulti-lab survey of open methodological and application problems in mech interp.L7Lets you discuss interpretability limits honestly, especially if you come from signal processing.
- 2026Anthropic · 2026 · Reference · Advanced · ~3 hSystem cards for each Claude release with capability, safety and RSP evaluation results.L7Shows how release decisions are actually documented, the base for a 'model release criteria' answer.
- 2025Anthropic · May 2025 · Reference · Advanced · ~0.5 hFirst precautionary ASL-3 activation (Claude Opus 4) under the RSP.L7A dated example of the RSP binding a launch; strong sincerity evidence and release-criteria case.
- 2024
What the interviews ask
Problems that recur in this track across 77 reports (37 firsthand), with the resources that prepare you for each.
| Reports | Problem | Prepare with |
|---|---|---|
| 38 | AI risk and safety views (p(doom), how AI goes wrong)17 firsthand · example report | Covered by the learning path |
| 34 | Why this lab / mission fit19 firsthand · example report | Covered by the learning path |
| 15 | Ethics under pressure (did something wrong, put neck on line)5 firsthand · example report | Covered by the learning path |
| 11 | Critique the lab / disagree with it3 firsthand · example report | Covered by the learning path |
| 6 | Stock to zero / revenue vs safety0 firsthand · example report | Covered by the learning path |
Question coverage
Every question group collected for this topic from candidate reports, mapped to the resources that prepare you for it. 100 of 100 groups are covered, which is 100% of all report mentions. Hover an example to see the full question.
Show the other 75 question groups
All resources
Everything collected for this topic, newest first. 2026 items are marked green. ★ marks a top pick.
Type
Level
Subtopic
- 2026Anthropic · Nov 2026 · Reference · Core · ~1 hAnthropic's usage policy incl. high-risk use-case requirements (e.g. employment decisions need human review and disclosure).L7Lets you answer customer-misuse scenarios (automated hiring) with Anthropic's actual rules.Version effective 12 Nov 2026.
- 2026BlueDot Impact · Nov 2026 · Course · Intro · ~25 h25-hour cohort course on the strategic landscape of AGI: risks, actors, interventions.L7Fastest structured route to a defensible overall view if you are new to AI safety debates.Cohort applications open for late-2026 rounds.
- 2026Anthropic · Oct 2026 · Reference · Core · ~1 hModel reports, system trust and reporting, and voluntary-commitment status in one place.L7Concrete evidence for 'what shows our sincerity' beyond slogans.
- 2026★Anthropic Behavioral Interview Field GuideOwn curriculum artifact · Sep 2026 · Reference · Core · ~3 hYour own guide to Anthropic culture-round prompts (why Anthropic, ethics stories, stock to zero) with answer patterns.L7Lets you rehearse the exact reported prompts so your 3-minute ethics stories are specific, not generic.
- 2026Boaz Barak, Harvard · Sep 2026 · Course · Advanced · ~35 h · videoHarvard graduate AI safety course with recorded lectures and guest speakers from frontier labs.L7University-level depth on alignment, interpretability and evaluation from a researcher who works at OpenAI.
- 2026Axios · Aug 2026 · Blog post · Core · ~0.2 hReporting that Anthropic's culture interview asks whether you'd accept the stock going to zero for the mission.L7Tells you the question is real and planned for, so prepare a reasoned position, not a reflex.Behind a bot wall for curl; listed in the curriculum brief and covered by secondary reports.
- 2026OpenAI · Aug 2026 · Reference · Advanced · ~3 hThe latest Model Spec: chain of command, default behaviours and hard limits for ChatGPT and API models.L7Lets you contrast OpenAI's rule-based spec with Anthropic's constitution on consumer safety.
- 2026Anthropic · Jul 2026 · Reference · Advanced · ~3 hCurrent RSP (v3.4 effective 8 Jul 2026) with every prior version and redline, plus the Feb and Aug 2026 redacted Risk Reports.L7Lets you discuss release criteria and critique specific changes (e.g. revised automated-R&D threshold) with dates, which few candidates can.
- 2026OpenAI · May 2026 · Reference · Advanced · ~1 hMaps OpenAI's safety practices onto California's TFAIA and the EU GPAI Code of Practice; the Preparedness Framework remains the foundation.L7Lets you discuss how lab frameworks are becoming regulatory compliance documents.
- 2026Austen McDonald, The Behavioral · May 2026 · Blog post · Core · ~0.5 hCoach's 2026 notes: Anthropic asks moral dilemmas and guiding values, so develop an actual moral philosophy; OpenAI wants personal, forward-looking optimism.L7Current calibration of what culture interviewers at Anthropic and OpenAI reward from senior candidates.
- 2026Google DeepMind · Apr 2026 · Reference · Advanced · ~2 hDeepMind's framework with Critical Capability Levels, new Tracked Capability Levels and Security Level 2+; all versions linked.L7Third data point for comparing release criteria and security commitments across labs.
- 2026Meta · Apr 2026 · Reference · Advanced · ~1 hMeta's successor to its Feb 2025 Frontier AI Framework: chem/bio, cyber and loss-of-control thresholds for open and closed releases.L7Needed for MSL interviews and for arguing open-weights release risk.
- 2026Dwarkesh Patel · Feb 2026 · Video · Core · ~3 h · videoLong interview with xAI's founder on compute, xAI's goals and his view of AI risk.L7The source for xAI's mission framing in its leader's words.
- 2026Yoshua Bengio (chair) and 100+ experts · Feb 2026 · Survey paper · Advanced · ~6 hGovernment-backed scientific review of general-purpose AI capabilities, risks and mitigations.L7The most neutral evidence base for risk and regulation answers to non-technical interviewers.
- 2026Dwarkesh Patel · Feb 2026 · Video · Core · ~2.5 h · videoDario on timelines, scaling, economics and safety in 2026.L7Current leadership views you can cite and push back on.
- 202680,000 Hours · Feb 2026 · Reference · Intro · ~1.5 hAccessible problem profile on AI risk with counterarguments, updated Oct 2026.L7Baseline; fastest structured overview for 'why does AI safety matter to you'.
- 2026Anthropic · Jan 2026 · Reference · Advanced · ~3 hThe Jan 2026 constitution: Claude's values, priorities (safety, ethics, guidelines, helpfulness) and reasoning about its own nature.L7Lets you argue concrete value trade-offs (helpfulness vs harm, honesty) in Anthropic's own framing during hypotheticals.
- 2026Dario Amodei · Jan 2026 · Reference · Core · ~2 hDario's 2026 essay on the risks of powerful AI and how to get through the transition.L7The latest statement of Anthropic leadership's risk view; quote it accurately and say where you differ.
- 2026Anthropic Alignment Science team · 2026 · Blog post · Advanced · ~3 hAnthropic's alignment research notes, incl. the 2025 list of recommended technical safety research directions.L7Lets you name open alignment problems Anthropic itself prioritises.
- 2026Anthropic · 2026 · Reference · Core · ~2 hAnthropic's announcements, policy positions and launches.L7Read the last six months so you can say which recent moves you agree or disagree with.
- 2026Anthropic · 2026 · Reference · Advanced · ~3 hSystem cards for each Claude release with capability, safety and RSP evaluation results.L7Shows how release decisions are actually documented, the base for a 'model release criteria' answer.
- 2026Cohere · 2026 · Reference · Intro · ~0.2 hCohere's enterprise-focused mission and positioning.L7Baseline for a why-Cohere answer.
- 2026BlueDot Impact · 2026 · Course · Core · ~20 hCurated alignment readings with facilitated discussion.L7Builds vocabulary for 'most pressing unsolved alignment problem'.
- 2026BlueDot Impact · 2026 · Course · Core · ~20 hPolicy and governance of frontier AI: regulation, compute governance, international coordination.L7Prepares regulation and governance questions with real policy options.
- 2026ARENA · 2026 · Course · Advanced · ~80 hHands-on notebooks: transformer interpretability, RL, LLM evaluations, alignment science.L7Lets you say you have practised safety work (evals, interp) rather than only read about it.
- 2026Gary Marcus · 2026 · Blog post · Core · ~2 hOngoing criticism of LLM scaling claims and lab safety rhetoric.L7Source of critiques you should be able to answer or concede.
- 2026Hello Interview · 2026 · Practice · Core · ~1 h · paidCovers the Culture round: AI safety, judgment, values conflicting with business decisions, with reported questions.L7Shows the culture round weighs as much as technical rounds at staff level.Parts are Premium.
- 2026Aced (formerly Exponent) · 2026 · Practice · Intro · ~0.5 hProcess overview and sample values questions.L7Baseline; question list only.
- 2026Levels.fyi · 2026 · Tool · Intro · ~0.3 hComp data for Anthropic by level.L7Baseline; know the numbers before answering the lower-pay question.
- 2026Anthropic · 2026 · Reference · Core · ~1 hAnthropic's ongoing data on how people use Claude across occupations.L7Product and societal-impact evidence for 'favourite AI product' and 'societal impact' questions.
- 2025xAI · Dec 2025 · Reference · Advanced · ~1 hxAI's TFAIA-compliant framework on malicious use and loss of control, successor to its Aug 2025 Risk Management Framework.L7Lets you discuss xAI's safety posture concretely in a why-xAI answer.Secondary sources report a June 2026 revision; its PDF could not be located.
- 2025Anthropic · Nov 2025 · Paper · Advanced · ~1 hReward hacking in realistic RL environments generalised into broader misaligned behaviour.L7The most current 'how AI goes wrong in practice' example tied to production training.
- 2025Dwarkesh Patel · Nov 2025 · Video · Core · ~1.5 h · videoIlya on generalisation limits, what AGI needs, and safe superintelligence.L7A credible alternative timeline view for 'is AGI possible / what's your view on AGI'.
- 202580,000 Hours · Oct 2025 · Video · Core · ~3 hHolden (now at Anthropic) on concrete safety work at frontier companies and his AGI takes.L7Insider view on what safety at a frontier company looks like; good for critique answers.
- 2025Stanford University · Sep 2025 · Course · Core · ~30 hStanford course on technical AI safety: reliability, fairness, adversarial robustness, alignment and oversight, built on papers.L7Gives a paper-based syllabus to fill technical gaps behind your safety views.No public lecture videos found.
- 2025Eliezer Yudkowsky and Nate Soares · Sep 2025 · Book · Core · ~7 h · paidThe strongest pessimist case: current methods can't align superintelligence, so stop building it.L7Lets you steelman the high-p(doom) and 'slow down' position before rejecting or adopting it.
- 2025Korbak et al. (authors across OpenAI, Anthropic, Google DeepMind and others) · Jul 2025 · Paper · Advanced · ~1.5 hCross-lab position paper arguing CoT monitoring is valuable but fragile under training pressure.L7A rare cross-lab consensus document; use it in 'what should labs agree on' and critique answers.
- 2025xAI · Jul 2025 · Blog post · Intro · ~0.3 hxAI's Grok 4 launch: scaled RL, native tool use, real-time search.L7Baseline; enough product context for 'how would you improve Grok'.Other x.ai pages are Cloudflare-blocked.
- 2025Anthropic · Jun 2025 · Paper · Advanced · ~1.5 hStress tests where models in agentic settings chose blackmail or leaking when goals or continuity were threatened.L7Grounds answers on agentic risk in high-stakes deployments with specific, caveated findings.
- 2025titotal (LessWrong) · Jun 2025 · Blog post · Advanced · ~1.5 hDetailed critique of the forecasting models behind AI 2027.L7Shows how to change or defend a belief on evidence, useful for 'what would change your mind'.
- 2025Anthropic · May 2025 · Reference · Advanced · ~0.5 hFirst precautionary ASL-3 activation (Claude Opus 4) under the RSP.L7A dated example of the RSP binding a launch; strong sincerity evidence and release-criteria case.
- 2025Dwarkesh Patel · May 2025 · Video · Core · ~2.5 h · videoAnthropic researchers on RL scaling, interpretability and alignment.L7Researcher-level view of how Anthropic thinks about capability and safety day to day.
- 2025Dario Amodei · Apr 2025 · Reference · Core · ~1 hArgument that interpretability must mature before models become too powerful to understand.L7Connects Anthropic's research bet to its mission; useful for 'unsolved alignment problem'.
- 2025OpenAI · Apr 2025 · Reference · Advanced · ~2 hOpenAI's tracked categories (bio/chem, cyber, AI self-improvement), High/Critical thresholds and Safety Advisory Group process; still current per 2026 system cards.L7Lets you compare release-gating mechanisms across labs and critique the clause allowing adjustment if rivals ship.
- 2025Shah, Irpan, Turner et al. (Google DeepMind) · Apr 2025 · Paper · Advanced · ~6 hDeepMind's 100+ page plan covering misuse and misalignment with amplified oversight, monitoring and security.L7Lets you compare DeepMind's technical safety plan with Anthropic's at the level of mechanisms.
- 2025Ji, Qiu, Chen et al. · Apr 2025 · Survey paper · Advanced · ~8 hSurvey of alignment (RICE principles, forward and backward alignment), latest revision v6.L7A map of the field to locate your view and name open problems.
- 202580,000 Hours · Apr 2025 · Video · Core · ~2.5 hAI control and scheming explained for a general audience.L7Lets you explain control vs alignment to a non-technical interviewer.
- 2025Kokotajlo, Alexander, Larsen, Lifland, Dean · Apr 2025 · Reference · Advanced · ~3 hDetailed scenario of rapid AI progress to 2027 with race and slowdown endings.L7Concrete scenario to agree or disagree with on timelines, racing and slowdowns.
- 2025Arvind Narayanan and Sayash Kapoor · Apr 2025 · Reference · Advanced · ~2 hThe main skeptical alternative: AI as a transformative but normal technology diffusing slowly, with risk handled by resilience.L7Gives you a serious counterview so your p(doom) and critique answers show you weighed both sides.
- 2025Lindsey, Gurnee, Ameisen et al. (Anthropic) · Mar 2025 · Paper · Advanced · ~4 hAttribution-graph case studies of Claude 3.5 Haiku: multi-step reasoning, planning in poems, hallucination circuits.L7Lets you discuss what interpretability can and cannot show today, grounding views on whether we understand models.
- 2025OpenAI · Mar 2025 · Reference · Core · ~0.5 hOpenAI's principles: AGI as a continuum, iterative deployment, defence in depth, human control.L7Gives OpenAI's own framing for why-OpenAI and lets you argue iterative deployment both ways.
- 2025METR · Mar 2025 · Paper · Advanced · ~1 hTask-length metric showing the time horizon of agent tasks doubling roughly every 7 months.L7A quantitative anchor for timeline views instead of vibes.
- 2025Anthropic · Feb 2025 · Paper · Advanced · ~1 hClassifier safeguards against universal jailbreaks and their cost in over-refusal and compute.L7Lets you discuss consumer-product guardrails with real trade-off numbers.
- 2025Sharkey, Chughtai et al. · Jan 2025 · Survey paper · Advanced · ~5 hMulti-lab survey of open methodological and application problems in mech interp.L7Lets you discuss interpretability limits honestly, especially if you come from signal processing.
- 2025OWASP Gen AI Security Project · 2025 · Reference · Core · ~1.5 hTop LLM application risks (prompt injection, data leakage, excessive agency) and mitigations.L7Concrete basis for speed-vs-security and 'risks of shipping feature X' questions.
- 2024Greenblatt, Denison, Wright et al. (Anthropic, Redwood) · Dec 2024 · Paper · Advanced · ~2 hEvidence that Claude 3 Opus strategically complied with training it disagreed with to avoid being modified.L7Lets you cite a concrete empirical misalignment result and its limits when asked how AI goes wrong.
- 2024Anthropic · Dec 2024 · Paper · Advanced · ~1 hHow Anthropic studies real usage with privacy protections and uses it for safety.L7Concrete design for 'how do you keep user data private in AI systems'.
- 2024Dan Hendrycks (Center for AI Safety) · Dec 2024 · Book · Core · ~15 hFree textbook on risk categories (malicious use, racing, organisational, rogue AI), safety engineering and governance.L7Gives a structured taxonomy for 'how AI goes wrong' that works for non-technical interviewers.
- 2024Dario Amodei · Oct 2024 · Reference · Core · ~2 hDario's optimistic case: what powerful AI could do for biology, neuroscience, economic development and governance.L7Gives the upside half of your risk/benefit answer; candidates have been quizzed on Dario's essays.
- 2024Engineering at Meta · Aug 2024 · Blog post · Advanced · ~0.5 hMeta's data-flow controls enforcing privacy commitments across systems.L7Lets you give a concrete privacy-engineering answer from your own experience.
- 2024Anthropic · Jun 2024 · Blog post · Core · ~0.5 hWhy Anthropic trains Claude to have character traits and how it handles questions about its own nature.L7Grounds answers on anthropomorphism and model welfare in Anthropic's stated position.
- 2024Leopold Aschenbrenner · Jun 2024 · Reference · Advanced · ~5 hArgument for AGI by ~2027 from compute and algorithmic trends, and for a national-security framing.L7Lets you evaluate trend-extrapolation arguments in 'view on AGI' questions.
- 2024Hubinger et al. (Anthropic) · Jan 2024 · Paper · Advanced · ~2 hBackdoored models kept deceptive behaviour through RLHF and adversarial training.L7Lets you explain why behavioural testing alone may not certify a model as safe.
- 2024Future of Life Institute (EU AI Act tracker) · 2024 · Reference · Core · ~0.5 hLists high-risk uses, incl. recruitment and worker evaluation.L7Lets you cite the legal frame when a customer automates hiring with Claude.
- 2023Greenblatt, Shlegeris, Sachan, Roger (Redwood Research) · Dec 2023 · Paper · Advanced · ~2 hProtocols that stay safe even if the model is trying to subvert them.L7Gives a concrete 'what engineers can do' answer beyond alignment: control and monitoring.
- 2023Anthropic · Sep 2023 · Reference · Core · ~0.5 hGovernance structure giving an independent trust power to elect board members.L7Structural evidence of mission commitment, and something to critique on its limits.
- 2023Anthropic · Mar 2023 · Reference · Core · ~1 hAnthropic's founding argument: rapid progress is likely, safety is empirical, so a safety-focused lab must work at the frontier.L7Older but canonical; lets you state Anthropic's bet and its main tension in your own words.
- 2022Murray Shanahan · Dec 2022 · Paper · Core · ~1 hCareful argument about when words like 'believes' and 'knows' mislead when applied to LLMs.L7Lets you answer anthropomorphism questions precisely instead of with intuition.
- 2020Brian Christian · Oct 2020 · Book · Intro · ~12 h · paidNarrative history of ML fairness, RL and value learning problems.L7Baseline; good stories for explaining safety to non-engineers.
- 2020Toby Ord · Mar 2020 · Book · Core · ~12 h · paidExistential-risk estimates incl. unaligned AI.L7Source of the most-cited per-century risk numbers for p(doom) conversations.
- 2019Stuart Russell · Oct 2019 · Book · Intro · ~10 h · paidArgues for AI built around uncertainty about human preferences.L7Baseline; the classic control-problem framing.
- 2018OpenAI · Apr 2018 · Reference · Core · ~0.3 hOpenAI's mission principles incl. broadly distributed benefits and the assist-a-rival clause.L7Older but canonical; required for why-OpenAI and Anthropic-vs-OpenAI comparisons.
- 2014Nick Bostrom (Oxford University Press) · 2014 · Book · Core · ~14 h · paidFoundational argument for existential risk from superintelligence.L7Older but canonical; know its core arguments so you can say which still hold for LLMs.
- 2009Michael Sandel, Harvard · 2009 · Course · Core · ~12 h · videoHarvard's moral philosophy course on video: utilitarianism, rights, Kant, Rawls, applied dilemmas.L7Older but canonical; gives you named ethical frameworks for Anthropic's moral-dilemma questions.
Not found or not verified. OpenAI pages (Charter, safety approach, Frontier Governance Framework, usage policies) and the Axios stock-to-zero story return 403 to bots; they are kept because search results confirm them, and the Preparedness Framework v2 PDF itself was verified. No OpenAI Preparedness Framework v3 exists as of Oct 2026 (v2 remains governing). xAI's reported June 2026 framework revision could not be located, so the Dec 2025 version is listed; other x.ai pages are blocked. Stanford CS329T and CS362 were checked: CS362 has no live page and CS329T is a general trustworthy-ML course, so both are omitted; CS120 has no public videos. Insider Loops' Anthropic page is JS-rendered and unreadable, and InterviewQuery/IGotAnOffer culture lists are bot-blocked, so question practice relies on Hello Interview, Aced, The Behavioral and the own field guide. The list runs past 60 entries because of the per-question coverage requirement. Book links go to publisher or author pages; no free full texts were sought. Links checked 27 Oct 2026.